A multi-dimensional fusion method for cross-departmental enterprise credit information knowledge graph

Abstract

The present application relates to a multi-dimensional fusion method for cross-departmental enterprise credit information knowledge graphs, which includes collecting multi-dimensional data from enterprise credit information knowledge graphs of two different departments and indexing the multi-dimensional data; performing feature extraction on the indexed multi-dimensional data in the two departments to obtain multi-dimensional data embedding representations in the source knowledge graph and the target knowledge graph, respectively; calculating similarity value matrices of multi-dimensional data embedding representations of entities in the source knowledge graph and the target knowledge graph; fusing the similarity value matrices of corresponding entity pairs between the source knowledge graph and the target knowledge graph to obtain a fused similarity representation vector for each entity pair; inputting the fused similarity representation vector of each entity pair into a multi-layer perceptron to obtain a final similarity score for each entity pair; and performing alignment and fusion based on the entity pairs with the highest final similarity scores to obtain a cross-departmental enterprise credit information fused knowledge graph.

Description

Technical Field

[0001] The present application relates to the technical field of knowledge graph fusion, and in particular to a multi-dimensional fusion method of cross-departmental enterprise credit information knowledge graphs. Background Art

[0002] At present, the knowledge graph fusion methods for corporate credit have the following main shortcomings: 1) There are a large number of entity overlaps in cross-departmental corporate credit information. For example, the same enterprise recorded in different parts may use different names or logos. The existing entity overlap detection methods are inefficient and inaccurate when dealing with such diversity and complexity; 2) Most of the current knowledge graph fusion technologies focus on the fusion of a single dimension (such as text), and lack the ability to integrate other dimensions such as images and time. In the cross-departmental corporate credit information fusion, the company's logo (image dimension), establishment time (time dimension) and other information are key data. Existing technologies are difficult to integrate multi-dimensional information for effective fusion, which affects the comprehensiveness and accuracy of the knowledge graph. Summary of the invention

[0003] Based on this, it is necessary to provide a multi-dimensional fusion method for cross-departmental enterprise credit information knowledge graph, which includes:

[0004] S1: Collect multi-dimensional data in the enterprise credit information knowledge graphs of two different departments and index the multi-dimensional data;

[0005] S2: Extract features from the indexed multi-dimensional data in the two departments to obtain the embedded representation of the multi-dimensional data in the source knowledge graph and the target knowledge graph respectively;

[0006] S3: Calculate the similarity value matrix of the multi-dimensional data embedding representation of each entity in the source knowledge graph and the target knowledge graph;

[0007] S4: Fuse the similarity value matrices of the corresponding entity pairs between the source knowledge graph and the target knowledge graph to obtain a fused similarity representation vector for each entity pair;

[0008] S5: inputting the fused similarity representation vector of each entity pair into a multi-layer perceptron to obtain a final similarity score of each entity pair;

[0009] S6: Align and fuse the entity pairs with the highest final similarity scores to obtain a cross-departmental enterprise credit information fusion knowledge graph.

[0010] Preferably, the multi-dimensional data includes the enterprise entity name, the enterprise logo image and the enterprise establishment time; the enterprise logo image and the enterprise establishment time are attributes of the entity; and indexing the multi-dimensional data includes:

[0011] The enterprise entity name, enterprise logo image and enterprise establishment time are generated into sequences according to the different enterprise entity names. The corresponding index value of ; where n represents the total number of enterprises; z It means any one of the following: the name of the corporate entity, the corporate logo image, and the date of establishment of the company.

[0012] Preferably, before feature extraction, the method further includes:

[0013] Construct positive seed entity sets and negative seed entity sets; the positive seed entity set is recorded as: ; The negative seed entity set is recorded as: ;in, Represents the first i Entity; Represents the first i Entity; Represents the first j Entity;

[0014] Based on the positive seed entity set and the negative seed entity set, the enterprise entity alignment probability is calculated using contrastive learning. The calculation formula is:

[0015] ;

[0016] ;

[0017] ;

[0018] in, represents the probability of enterprise entity alignment between two knowledge graphs; T represents transposition; represents the temperature coefficient; Indicates the similarity of entity pairs;

[0019] The contrastive learning loss is calculated based on the alignment probability of the enterprise entities, and the calculation formula is:

[0020] ;

[0021] in, represents contrastive learning loss;

[0022] The pre-trained BERT model and the VGG16 pre-trained vision model are adjusted by minimizing the contrastive learning loss.

[0023] Preferably, the feature extraction includes:

[0024] For the multi-dimensional data after index processing in any department,

[0025] Extract features of the enterprise entity name according to the adjusted pre-trained BERT model to obtain an embedded representation of the enterprise entity name information;

[0026] Extracting features of the corporate logo image according to the adjusted VGG16 pre-trained visual model to obtain an embedded representation of the corporate logo image information;

[0027] The feature extraction of the enterprise establishment time is performed according to the Time2Vec model to obtain an embedded representation of the enterprise establishment time information.

[0028] Preferably, the feature extraction of the enterprise establishment time according to the Time2Vec model includes the Time2Vec model mapping the enterprise establishment time into an embedded representation of the enterprise establishment time information through a mapping function; the mapping function expression is:

[0029] ;

[0030] in, Indicates that the company establishment time information is embedded in the representation x elements; represents a periodic sinusoidal activation function; represents the frequency in the periodic sinusoidal activation function; represents the phase shift in the periodic sinusoidal activation function; d Indicates the dimension of the embedded representation of multi-dimensional data; Represents a time scalar.

[0031] Preferably, the similarity value matrix of the multi-dimensional data embedding representation of each entity in the source knowledge graph and the target knowledge graph is calculated to include:

[0032] Combining all the business entity name information embedding representations in any graph into a first embedding representation matrix;

[0033] Combining all the corporate logo image information embedding representations in any one of the graphs into a second embedding representation matrix;

[0034] Combining all the enterprise establishment time information embedding representations in any graph into a third embedding representation matrix;

[0035] Calculate the similarity value matrix of the enterprise entity name based on the first embedding representation matrix in the source knowledge graph and the target knowledge graph;

[0036] Calculate the similarity value matrix of the corporate logo image based on the second embedding representation matrix in the source knowledge graph and the target knowledge graph;

[0037] The similarity value matrix of enterprise establishment time is calculated based on the third embedding representation matrix in the source knowledge graph and the target knowledge graph.

[0038] Preferably, the similarity value matrix calculation formula is:

[0039] ;

[0040] in, Represents any similarity value matrix; An arbitrary embedding representation matrix representing the source knowledge graph; Represents the target knowledge graph with The corresponding embedding representation matrix; Represents the embedding representation matrix No. l dimension; Represents the embedding representation matrix No. l dimension; d Indicates the dimension of the embedding representation of multi-dimensional data.

[0041] Preferably, the calculation formula for the fusion similarity representation vector of the entity pair is:

[0042] ;

[0043] in, Represents entity pair The fusion similarity representation vector of Represents entities in the source knowledge graph; Represents entities in the target knowledge graph; A matrix of similarity values ​​representing business entity names; Represents the similarity value matrix of corporate logo images; Represents the similarity value matrix of the establishment time of the enterprise; Represents a concatenation operation.

[0044] Preferably, the calculation formula for the final similarity score of the entity pair is:

[0045] ;

[0046] in, Represents entity pair The final similarity score of Represents entities in the source knowledge graph; Represents entities in the target knowledge graph; Represents a multilayer perceptron.

[0047] Preferably, the entity pairs with the highest final similarity scores are aligned and fused to obtain a cross-departmental enterprise credit information fusion knowledge graph, including:

[0048] Align and fuse the enterprise entity names corresponding to the entity pair with the highest final similarity score; record the entity pair with the highest final similarity score as the aligned entity pair, record the entity from the source knowledge graph in the aligned entity pair as the first entity, and record the entity from the target knowledge graph in the aligned entity pair as the second entity;

[0049] For the relationship associated with the first entity pair, after the entities are aligned, connecting the relationship corresponding to the second entity pair to the first entity;

[0050] For relations not associated with the first entity pair, after entity alignment, the second entity and its corresponding relations are added to the source knowledge graph;

[0051] The attributes of each entity are aligned and fused according to the enterprise logo image and the index value corresponding to the establishment time of each entity in the source knowledge graph and the target knowledge graph. After the entity is aligned, only the attributes of the entity in the source knowledge graph are retained.

[0052] Beneficial effects: This method significantly improves the accuracy of cross-departmental knowledge graph fusion through the comprehensive utilization of multi-dimensional data, can provide richer corporate portraits, and solves the shortcomings of traditional methods in dealing with diversity and complexity; by adopting an efficient multi-dimensional data similarity value matrix and fusion representation calculation mechanism, it significantly improves the efficiency of knowledge graph fusion and can quickly process large-scale data; the constructed cross-departmental corporate credit information fusion knowledge graph can provide reliable data support for various applications such as corporate credit evaluation and risk warning, thereby enhancing the practical value of the data. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0054] Figure 1 This is a flow chart of the multi-dimensional integration method of the cross-departmental enterprise credit information knowledge graph according to an embodiment of the present application. DETAILED DESCRIPTION

[0055] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are described in detail below in conjunction with the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present application. However, the present application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present application, so the present application is not limited by the specific embodiments disclosed below.

[0056] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of this application, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0057] like Figure 1 As shown, this embodiment provides a multi-dimensional fusion method for cross-departmental enterprise credit information knowledge graph, the method comprising:

[0058] S1: Collect multi-dimensional data from the enterprise credit information knowledge graphs of two different departments and index the multi-dimensional data.

[0059] Specifically, the multi-dimensional data includes an enterprise entity name, an enterprise logo image, and an enterprise establishment time; the enterprise logo image and the enterprise establishment time are attributes of the entity; and indexing the multi-dimensional data includes:

[0060] The enterprise entity name, enterprise logo image and enterprise establishment time are generated into sequences according to the different enterprise entity names. The corresponding index value of ; where n represents the total number of enterprises; z It means any one of the following: the name of the corporate entity, the corporate logo image, and the date of establishment of the company.

[0061] S2: Perform feature extraction on the indexed multi-dimensional data in the two departments to obtain the embedded representation of the multi-dimensional data in the source knowledge graph and the target knowledge graph respectively.

[0062] In this example, before feature extraction, the following steps are also performed:

[0063] Construct positive seed entity sets and negative seed entity sets; the positive seed entity set is recorded as: ; The negative seed entity set is recorded as: ;in, Represents the first i Entity; Represents the first i Entity; Represents the first j Entity;

[0064] Based on the positive seed entity set and the negative seed entity set, the enterprise entity alignment probability is calculated using contrastive learning. The calculation formula is:

[0065] ;

[0066] ;

[0067] ;

[0068] in, represents the probability of enterprise entity alignment between two knowledge graphs; T represents transposition; represents the temperature coefficient; Indicates the similarity of entity pairs;

[0069] The contrastive learning loss is calculated based on the enterprise entity alignment probability, and the calculation formula is:

[0070] ;

[0071] in, represents contrastive learning loss;

[0072] The pre-trained BERT model and the VGG16 pre-trained vision model are adjusted by minimizing the contrastive learning loss.

[0073] Furthermore, the feature extraction includes:

[0074] For the multi-dimensional data after index processing in any department,

[0075] Extract features of the enterprise entity name according to the adjusted pre-trained BERT model to obtain an embedded representation of the enterprise entity name information;

[0076] Extracting features of the corporate logo image according to the adjusted VGG16 pre-trained visual model to obtain an embedded representation of the corporate logo image information;

[0077] The time of establishment of the enterprise is feature extracted according to the Time2Vec model to obtain an embedded representation of the time of establishment information of the enterprise. The Time2Vec model maps the time of establishment of the enterprise to an embedded representation of the time of establishment information of the enterprise through a mapping function; the mapping function expression is:

[0078] ;

[0079] in, Indicates that the company establishment time information is embedded in the representation x elements; represents a periodic sinusoidal activation function; represents the frequency in the periodic sinusoidal activation function; represents the phase shift in the periodic sinusoidal activation function; d Indicates the dimension of the embedded representation of multi-dimensional data; Represents a given time scalar.

[0080] S3: Calculate the similarity value matrix of the multi-dimensional data embedding representation of each entity in the source knowledge graph and the target knowledge graph.

[0081] Specifically, all the enterprise entity name information in any graph are embedded and represented and combined into a first embedding representation matrix;

[0082] Combining all the corporate logo image information embedding representations in any one of the graphs into a second embedding representation matrix;

[0083] Combining all the enterprise establishment time information embedding representations in any graph into a third embedding representation matrix;

[0084] Calculate the similarity value matrix of the enterprise entity name based on the first embedding representation matrix in the source knowledge graph and the target knowledge graph;

[0085] Calculate the similarity value matrix of the corporate logo image based on the second embedding representation matrix in the source knowledge graph and the target knowledge graph;

[0086] The similarity value matrix of enterprise establishment time is calculated based on the third embedding representation matrix in the source knowledge graph and the target knowledge graph.

[0087] Furthermore, considering that cosine similarity is used to calculate the angle between two vectors, that is, it measures the degree of similarity between two vectors in direction, it takes a value between -1 and 1, which represents the degree of directional similarity between vectors. 1 represents exact similarity, 0 represents orthogonality, and -1 represents complete oppositeness. The larger the value, the better. Euclidean distance is a measure of the straight-line distance between two points in Euclidean space, which is directly related to the position of the vector. The smaller the value, the better. Therefore, considering the position and direction between the embedding representation matrices at the same time can more accurately calculate their similarity. Therefore, the similarity value matrix calculation formula is:

[0088] ;

[0089] in, Represents any similarity value matrix; An arbitrary embedding representation matrix representing the source knowledge graph; Represents the target knowledge graph with The corresponding embedding representation matrix; Represents the embedding representation matrix No. l dimension; Represents the embedding representation matrix No. l dimension; d Indicates the dimension of the embedding representation of multi-dimensional data.

[0090] S4: Fuse the similarity value matrices of the corresponding entity pairs between the source knowledge graph and the target knowledge graph to obtain the fused similarity representation vector of each entity pair.

[0091] Specifically, the calculation formula for the fusion similarity representation vector of the entity pair is:

[0092] ;

[0093] in, Represents entity pair The fusion similarity representation vector of Represents entities in the source knowledge graph; Represents entities in the target knowledge graph; A matrix of similarity values ​​representing business entity names; Represents the similarity value matrix of corporate logo images; Represents the similarity value matrix of the establishment time of the enterprise; Represents a concatenation operation.

[0094] S5: Input the fused similarity representation vector of each entity pair into a multi-layer perceptron to obtain a final similarity score for each entity pair.

[0095] Specifically, the calculation formula for the final similarity score of the entity pair is:

[0096] ;

[0097] in, Represents entity pair The final similarity score of Represents entities in the source knowledge graph; Represents entities in the target knowledge graph; Represents a multilayer perceptron.

[0098] S6: Align and fuse the entity pairs with the highest final similarity scores to obtain a cross-departmental enterprise credit information fusion knowledge graph.

[0099] Specifically, the step includes:

[0100] Align and fuse the enterprise entity names corresponding to the entity pair with the highest final similarity score; record the entity pair with the highest final similarity score as the aligned entity pair, record the entity from the source knowledge graph in the aligned entity pair as the first entity, and record the entity from the target knowledge graph in the aligned entity pair as the second entity;

[0101] For the relationship associated with the first entity pair, after the entities are aligned, connecting the relationship corresponding to the second entity pair to the first entity;

[0102] For relations not associated with the first entity pair, after entity alignment, the second entity and its corresponding relations are added to the source knowledge graph;

[0103] The attributes of each entity are aligned and fused according to the enterprise logo image and the index value corresponding to the establishment time of each entity in the source knowledge graph and the target knowledge graph. After the entity is aligned, only the attributes of the entity in the source knowledge graph are retained.

[0104] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0105] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be construed as limiting the scope of the patent application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent application shall be subject to the attached claims.

Claims

1. A multi-dimensional fusion method for cross-departmental enterprise credit information knowledge graph, characterized in that: include: S1: Collect multi-dimensional data in the enterprise credit information knowledge graphs of two different departments and index the multi-dimensional data; The multi-dimensional data includes the name of the enterprise entity, the enterprise logo image and the time of establishment of the enterprise; The enterprise logo image and the establishment time of the enterprise are attributes of the entity; The indexing of the multi-dimensional data comprises: The enterprise entity name, enterprise logo image and enterprise establishment time are generated into sequences according to the different enterprise entity names. The corresponding index value of ; where n represents the total number of enterprises; z Any one of the following: the name of the corporate entity, the corporate logo image, and the date of establishment of the company; S2: Extract features from the indexed multi-dimensional data in the two departments to obtain the embedded representation of the multi-dimensional data in the source knowledge graph and the target knowledge graph respectively; Before feature extraction, it also includes: Construct positive seed entity sets and negative seed entity sets; the positive seed entity set is recorded as: ; The negative seed entity set is recorded as: ;in, Represents the first i Entity; Represents the first i Entity; Represents the first j Entity; Based on the positive seed entity set and the negative seed entity set, the enterprise entity alignment probability is calculated using contrastive learning. The calculation formula is: ； ； ； in, represents the probability of enterprise entity alignment between two knowledge graphs; T represents transposition; represents the temperature coefficient; Indicates the similarity of entity pairs; The contrastive learning loss is calculated based on the alignment probability of the enterprise entities, and the calculation formula is: ； in, represents contrastive learning loss; Adjusting the pre-trained BERT model and the VGG16 pre-trained vision model by minimizing the contrastive learning loss; The feature extraction comprises: For the multi-dimensional data after index processing in any department, Extract features of the enterprise entity name according to the adjusted pre-trained BERT model to obtain an embedded representation of the enterprise entity name information; Extracting features of the corporate logo image according to the adjusted VGG16 pre-trained visual model to obtain an embedded representation of the corporate logo image information; Extract features of the enterprise establishment time according to the Time2Vec model to obtain an embedded representation of the enterprise establishment time information; The feature extraction of the enterprise establishment time according to the Time2Vec model includes the Time2Vec model mapping the enterprise establishment time into the embedded representation of the enterprise establishment time information through a mapping function; the mapping function expression is: ； in, Indicates that the company establishment time information is embedded in the representation x elements; represents a periodic sinusoidal activation function; represents the frequency in the periodic sinusoidal activation function; represents the phase shift in the periodic sinusoidal activation function; d Indicates the dimension of the embedded representation of multi-dimensional data; represents a time scalar; S3: Calculate the similarity value matrix of the multi-dimensional data embedding representation of each entity in the source knowledge graph and the target knowledge graph; S4: Fuse the similarity value matrices of the corresponding entity pairs between the source knowledge graph and the target knowledge graph to obtain a fused similarity representation vector for each entity pair; S5: inputting the fused similarity representation vector of each entity pair into a multi-layer perceptron to obtain a final similarity score of each entity pair; S6: Align and fuse the entity pairs with the highest final similarity scores to obtain a cross-departmental enterprise credit information fusion knowledge graph.

2. The multi-dimensional fusion method of cross-departmental enterprise credit information knowledge graph according to claim 1 is characterized in that: The similarity value matrix of the multi-dimensional data embedding representation of each entity in the computing source knowledge graph and the target knowledge graph includes: Combining all the business entity name information embedding representations in any graph into a first embedding representation matrix; Combining all the corporate logo image information embedding representations in any one of the graphs into a second embedding representation matrix; Combining all the enterprise establishment time information embedding representations in any graph into a third embedding representation matrix; Calculate the similarity value matrix of the enterprise entity name based on the first embedding representation matrix in the source knowledge graph and the target knowledge graph; Calculate the similarity value matrix of the corporate logo image based on the second embedding representation matrix in the source knowledge graph and the target knowledge graph; The similarity value matrix of enterprise establishment time is calculated based on the third embedding representation matrix in the source knowledge graph and the target knowledge graph.

3. The multi-dimensional fusion method of cross-departmental enterprise credit information knowledge graph according to claim 2 is characterized in that: The similarity value matrix calculation formula is: ； in, Represents any similarity value matrix; Any embedding representation matrix representing the source knowledge graph; Represents the target knowledge graph The corresponding embedding representation matrix; Represents the embedding representation matrix No. l dimension; Represents the embedding representation matrix No. l dimension; d Indicates the dimension of the embedding representation of multi-dimensional data.

4. The multi-dimensional fusion method of cross-departmental enterprise credit information knowledge graph according to claim 2 is characterized in that: The calculation formula of the fusion similarity representation vector of the entity pair is: ； in, Represents entity pair The fusion similarity representation vector of Represents entities in the source knowledge graph; Represents entities in the target knowledge graph; A matrix of similarity values ​​representing business entity names; Represents the similarity value matrix of corporate logo images; The similarity value matrix representing the establishment time of enterprises; Represents a concatenation operation.

5. The multi-dimensional fusion method of cross-departmental enterprise credit information knowledge graph according to claim 4 is characterized in that: The final similarity score of an entity pair is calculated as: ； in, Represents entity pair The final similarity score of Represents entities in the source knowledge graph; Represents entities in the target knowledge graph; Represents a multi-layer perceptron.

6. The multi-dimensional fusion method of cross-departmental enterprise credit information knowledge graph according to claim 5 is characterized in that: The entity pairs with the highest final similarity scores are aligned and fused to obtain a cross-departmental enterprise credit information fusion knowledge graph including: Align and fuse the enterprise entity names corresponding to the entity pair with the highest final similarity score; record the entity pair with the highest final similarity score as the aligned entity pair, record the entity from the source knowledge graph in the aligned entity pair as the first entity, and record the entity from the target knowledge graph in the aligned entity pair as the second entity; For the relationship associated with the first entity pair, after the entities are aligned, connecting the relationship corresponding to the second entity pair to the first entity; For relations not associated with the first entity pair, after entity alignment, the second entity and its corresponding relations are added to the source knowledge graph; The attributes of each entity are aligned and fused according to the enterprise logo image and the index value corresponding to the establishment time of each entity in the source knowledge graph and the target knowledge graph. After the entity is aligned, only the attributes of the entity in the source knowledge graph are retained.

Images